DeepSeek CEO Says Huawei Chips Could Catch Nvidia as Firm Prepares Training Shift

DeepSeek CEO Liang Wenfeng has made greater use of domestic AI chips a priority, telling investors that Huawei's processors could catch up with Nvidia's within a few years and that DeepSeek should adapt early to hardware outside the Nvidia ecosystem. Huawei could deliver a new batch of AI training chips to DeepSeek as early as the fourth quarter of 2026, though the scale of supply remains unclear, according to details from a closed-door investor meeting. The shift is significant because large-scale model training in China has still relied heavily on Nvidia hardware, even as U.S. export controls have restricted the PRC's ability to buy advanced computing chips and manufacture leading-edge chips. DeepSeek's earlier breakthrough was hardware-aware but still Nvidia-dependent: DeepSeek-V3 used 671B total parameters, 37B active parameters per token, 14.8T training tokens and 2.788M Nvidia H800 GPU hours. DeepSeek has since pushed further into Huawei compatibility. Its V4 preview listed DeepSeek-V4-Pro at 1.6T total parameters with 49B active parameters, and DeepSeek-V4-Flash at 284B total parameters with 13B active parameters. Liang told investors the company is now training a 2T-parameter model and plans an eventual 8T-parameter model. He also estimated that training a model comparable in scale to OpenAI's largest model under development would require about 50,000 Nvidia GB300 chips or about 200,000 Huawei Ascend 950 chips, a four-to-one chip-count gap that shows why the transition is both strategic and operational. Huawei's own chip roadmap gives DeepSeek a clearer path but not a guarantee. Huawei's Ascend roadmap says Ascend 950PR is optimized for prefill and recommendation workloads, while Ascend 950DT is optimized for decode and model training, with Ascend 950DT planned for the fourth quarter of 2026, 144 GB of memory, 4 TB/s memory bandwidth and 2 TB/s interconnect bandwidth. A newer roadmap accelerated Ascend 960DT to the first quarter of 2027 and Ascend 960PR to the third quarter of 2027, while Ascend 970 and Ascend 980 are slated for 2028 and 2029. But raw compute comparisons still favor Nvidia in many cases: published comparisons put Ascend 960DT at 2 FP8 PFLOPS versus Nvidia H200 at 4 FP8 PFLOPS, and Ascend 960PR at 8 FP4 PFLOPS versus Nvidia B300 at 15 to 20 NVFP4 PFLOPS. The reception among developers and infrastructure specialists is split. Supporters see the move as a practical route to cheaper domestic capacity, fewer bottlenecks and a less CUDA-dependent deployment stack. That view gained force after the April launch of DeepSeek V4, when Huawei said its chips were used in some of V4's training process and analysts described Ascend support as evidence that top Chinese models can run on Chinese hardware. Skeptics focus on training stability, interconnect performance, software maturity and supply. Earlier R2 training efforts reportedly ran into unstable performance, slower chip-to-chip connectivity and limits in Huawei's CANN software stack, pushing DeepSeek back to Nvidia for training while keeping Huawei hardware for inference. Huawei also faces capacity constraints, with next-generation Ascend 900-series accelerators expected to be focused on China rather than broad international sales. The next proof point is whether DeepSeek can move more pretraining work onto Huawei hardware without slowing its frontier-model cadence. For now, the Huawei shift looks less like an immediate Nvidia replacement than a staged attempt to build a domestic training-and-inference stack before the performance gap closes.